Recent progress of the operational dust prediction system in the Japan

Meteorological Agency

Akinori Ogi, Toshinori Aoyagi, Makoto Deushi Taichu Y. Tanaka, Keiya Yumimoto, Thomas T. Sekiyama, Takashi Maki

Japan Meteorological Agency

Meteorological Research Institute

12 July 2016 8th International Cooperative for Aerosol Prediction (ICAP)

@ NOAA Center for Weather and Climate Prediction

Outline

• Aeolian dust (Kosa) information to the public from JMA

• Verification of operational aerosol prediction, mainly focused on aeolian dust (Kosa) prediction

• Current development status and future planning of a new operational global aerosol forecast model for dust predictions by JMA

• Summary

Information on aeolian dust to the public

3

JMA also provides aeolian dust prediction results (GPV : GRIB2 format) for private weather services via the Japan Meteorological Business Support Center (JMBSC).

Aeolian dust observation

Aeolian dust advisory information (when required, Japanese only)

Statistics of aeolian dust

Basic knowledge about aeolian dust

Aeolian dust prediction

Blown up in arid area in the continent

Advected by the upper wind

Deposits in Japan

JMA has been providing aeolian dust information based on numerical forecasts and surface observations since January 2004.

Taklamakan Desert

Gobi Desert

Outline of the current operational global aerosol forecast model （MASINGAR mk-2）

Resolution TL159L40 Horizontal -110km, Vertical 40 layers (Surface – 0.4hPa)

Types of aerosols

10 bins of dust (0.2 - 20μm), 10 bins of sea salt (0.2 – 20μm), Sulfate, Organic carbon, Black carbon

Dust emission process

Depend on particle size, vegetation, surface condition (soil moisture, snow depth etc..) and surface wind speed

Dust deposition Process

Gravity (dry deposition), removal due to clouds and rain (wet deposition)

Dynamical model

MRI-AGCM3 (GSMUV)

Calculation interval

Once a day (12UTC initial)

Forecast period

5 days (120 hours)

Function of the surface friction velocity Dust emission flux

* No data assimilation of aerosol

Nudging Global Analysis

and forecast data

in JMA (GSM)

Output of calculation

result (every 3 hours)

In our daily operational prediction system, we’re combining the atmospheric general circulation model (GSMUV) with the global aerosol forecast model (MASINGAR mk-2). We updated the model from November 2014.

The MRI-ESM aims to improve the prediction of global warming. We apply this system to the daily aerosol prediction in JMA.

Verification of dust prediction - Statistical verification -

SYNOP reports at meteorological

observatories in Japan

Dust observation

（O）

Visibility becomes less than 10km because of aeolian dust.

Other phenomena (e.g. rainfall..) have not been seen

within an hour.

No dust observation

（X）

Aeolian dust that visibility becomes <10km has not been seen. Other phenomena have not also been seen within an

hour.

Unknown Other than those above.

(We cannot know whether the aeolian dust has been

observed because of the rainfall, etc..)

We calculate the statistics for dust predictions using SYNOP reports from meteorological observatories in Japan.

（Verification period：March-May 2010-2014, 00UTC-09UTC）

Dust forecast model

surface～1km conc.

Dust forecast

（F） ≧90μg/m3

No dust forecast

（X）

＜90μg/m3

• This threshold value is based on the past research results relating to the dust concentration and visibility. (Iwakura and Okada, 1999)

- Statistical verification -

Hit

RateMASINGAR MASINGAR mk-2

0 day 0.885 0.725

1 day 0.879 0.727

2 day 0.831 0.697

3 day 0.795 0.669

4 day 0.648 0.493

5 day 0.610 0.484

False

Alarm

Ratio

MASINGAR MASINGAR mk-2

0 day 0.643 0.531

1 day 0.642 0.528

2 day 0.650 0.542

3 day 0.659 0.548

4 day 0.701 0.633

5 day 0.703 0.645

Percent

CorrectMASINGAR MASINGAR mk-2

0 day 0.912 0.943

1 day 0.912 0.944

2 day 0.912 0.942

3 day 0.910 0.941

4 day 0.903 0.930

5 day 0.905 0.928

XOFXFO

FO

ScoreThreat

XOFO

FO

RateHit

FXFO

FX

Ratio Alarm False

XXFXXOFO

XXFO

CorrectPercent

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

0 1 2 3 4 5

Thre

at s

core

Forecast period (days)

Threat score for dust prediction in 2010-2014

MASINGAR

MASINGAR mk-2

- Quantitative verification - Model AOD forecast against satellite-based observation

(Average : Oct. 2014 – Oct. 2015)

- Quantitative verification - Model AOD forecast against satellite-based observation

According to the comparison with the MODIS AOD data, we have also seen a small positive bias in simulated AOD relative to MODIS AOD observations. The correlation coefficient is low in the summer and fall because of the uncertainty for smoke predictions in the operating system. So we are going to use the near real-time smoke data (GFAS daily fire products) to the operational aerosol prediction system.

- Quantitative verification - Surface AOD observation in JMA

JMA has been conducting AOD measurements using sun photometers at 3 WMO/GAW stations as a part of its environmental monitoring network.

Minamitorishima Yonagunijima → Ishigakijima (since Apr. 2016)

Ryori Precision Filter Radiometer (PFR)

Replacement of PFR to Sky Radiometer is now underway.

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

01-Mar-12 16-Mar-12 31-Mar-12 15-Apr-12 30-Apr-12 15-May-12 30-May-12

AO

D

Date

Ground-based AOD observations by the sun photometer vs. MASINGAR mk-2 model forecasts at Yonagunijima, Japan in 2012

model forecasts

observations

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

01-Mar-12 16-Mar-12 31-Mar-12 15-Apr-12 30-Apr-12 15-May-12 30-May-12

AO

D

Date

Ground-based AOD observations by the sun photometer vs. MASINGAR mk-2 model forecasts at Minamitorishima, Japan in 2012

model forecasts

observations

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

01-Mar-12 16-Mar-12 31-Mar-12 15-Apr-12 30-Apr-12 15-May-12 30-May-12

AO

D

Date

Ground-based AOD observations by the sun photometer vs. MASINGAR mk-2 model forecasts at Ryori, Japan in 2012

model forecasts

observations

- Quantitative verification - Model AOD forecast against ground-based observation

ME: 0.078 RMSE: 0.217 CC: 0.737

ME: 0.083 RMSE: 0.108 CC: 0.618

ME: 0.017 RMSE: 0.146 CC: 0.871

These results show a good correlation between ground-based AOD observations by the sun photometer and model forecasts. And there appears to be a small positive bias in these cases.

High-resolution global aerosol forecast model

TL159L40 (~110km) TL479L40 (~40km)

We have been developing a new version of the high-resolution global aerosol forecast model (MASINGAR mk-2 with TL479L40) and verifying the test data. Using GFAS daily fire products (thanks to ECMWF!), the high-resolution model simulates dense areas of smoke around northern part of Japan and Himawari-8 also represents these areas. And some bias correction methods are included in this version for reducing overestimation of AOD prediction.

Himawari-8 True Color Reproduction image 2016051906UTC

- Statistical verification -

XOFXFO

FO

ScoreThreat

XOFO

FO

RateHit

FXFO

FX

Ratio Alarm False

XXFXXOFO

XXFO

CorrectPercent

False

Alarm

Ratio

MASINGAR mk-2

TL159L40

MASINGAR mk-2

TL479L40

0 day 0.634 0.486

1 day 0.630 0.490

2 day 0.642 0.516

3 day 0.663 0.472

4 day 0.721 0.533

5 day 0.704 0.594

Percent

Correct

MASINGAR mk-2

TL159L40

MASINGAR mk-2

TL479L40

0 day 0.952 0.970

1 day 0.952 0.970

2 day 0.950 0.967

3 day 0.948 0.971

4 day 0.941 0.966

5 day 0.945 0.962

* A preliminary result

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

0 1 2 3 4 5

Thre

at s

core

Forecast period (days)

Threat score for dust prediction in 2013-2015

MASINGAR mk-2 TL159L40

MASINGAR mk-2 TL479L40

Hit

Rate

MASINGAR mk-2

TL159L40

MASINGAR mk-2

TL479L40

0 day 0.771 0.795

1 day 0.776 0.795

2 day 0.771 0.795

3 day 0.700 0.695

4 day 0.582 0.563

5 day 0.565 0.479

Summary • The current operational global aerosol forecast model

(MASINGAR mk-2 with TL159L40) has been used for dust predictions since Nov. 2014.

• The statistical verification results show the dust prediction is improved well in the current model and it can predict dust distributions better than the old one.

• The comparison between the AOD observations and the current model forecasts indicates a good performance although we have seen a small positive bias in the current version of the model.

• JMA has been developing a new version of the high-resolution forecast model (MASINGAR mk-2 with TL479L40) for the operational dust prediction system and we have evaluated its forecast accuracy.

• We are now testing this system for operational use.

That is all for my presentation. Thank you very much for your kind

attention!

Outline of the old operational global aerosol forecast model （MASINGAR）

Resolution T106L20 Horizontal -110km, Vertical 20 layers (Surface - 34hPa)

Type of aerosol

10 bins of dust (0.2 - 20μm)

Dust emission process

Depend on particle size, vegetation, surface condition (soil moisture, snow depth etc..) and surface wind speed

Dust deposition process

Gravity (dry deposition), removal due to clouds and rain (wet deposition)

Dynamical model

MRI/JMA98 (MJ98)

Calculation interval

Once a day (12UTC initial)

Forecast period

5 days (120 hours)

Dust emission flux

Surface wind Threshold of wind speed (> 6.5m/s)

* No data assimilation of dust The dust emission flux is proportional to the

cube of the wind speed.

[kg m-2 h-1]

30 March, 2012

12UTC ini

The part of dust calculation

・Advection ・Emission(Vegetation, soil moisture,

surface wind) ・Deposition (dry/wet) ・Dust particle : 0.2μm - 20 μm

Dynamical Model

MRI/JMA98(MJ98)

Nudging Global analysis

and forecast data

in JMA (GSM)

Output of calculation

result (every 3 hours)

Dust forecast model MASINGAR

Updates of the operational global aerosol forecast model

Old operational global dust forecast model

Current operational global aerosol forecast model

Global aerosol model MASINGAR (Tanaka et al., 2003) MASINGAR mk-2 (Tanaka et al., manuscript in preparation)

Dust emission Function of the wind speed (u10) F = C u10

2(u10 – ut)

Function of the surface friction velocity (Shao et al., 1996; Tanaka and Chiba, 2005)

Included aerosol species

Mineral dust Mineral dust, sulfate, BC, OA, sea salt

Resolution T106L20 (1.125°) TL159L40(1.125°) (in 2014) → TL479L40 (0.375°) (in 2017)

Atmospheric model MRI/JMA 98 AGCM (Shibata et al., 1998)

MRI-AGCM3 (Yukimoto et al., 2012)

Advection 3-dimensional semi-Lagrangian

Convective transport Arakawa-Schubert Yoshimura (Yoshimura et al.,2014)

Land surface model 3-layer Simple Biosphere HAL (Hosaka et al., manuscript in preparation)

Coupling of aerosol model with AGCM

Subroutine call in each time step Connected using SCUP library (Yoshimura and Yukimoto, 2008)

- Statistical verification - Visibility and meteorological conditions

• JMA operates 59 manned observational stations, which observe aeolian dust in terms of the visibility and meteorological conditions.

• The minimum visibility at each station is categorized in different colors on the JMA website.

• When the visibility becomes below 10 km, the station reports aeolian dust in SYNOP messages.

Map of stations observing aeolian dust Kosa or local sand/dust haze during the day

Distribution of stations observing aeolian dust

This observation is used for the validation of dust prediction with Threat Score (TS).

XXFXXOFO

XXFO

CorrectPercent

FO ： Forecast・Observation FX ： Forecast・No Observation

XOFXFO

FO

ScoreThreat

XO ： No Forecast・Observation XX ： No Forecast・No Observation

XOFO

FO

RateHit

FXFO

FX

Ratio Alarm False

It’s the fraction of observed events that are forecasted

correctly.

It’s the fraction of forecasts that are

correct.

It combines ‘Hit Rate’ and ‘False

Alarm Ratio’ into one score for low

frequency events.

It’s the fraction of forecasts that are wrong,

i.e., are false alarm.

- Statistical verification - How to calculate the statistics of dust prediction

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0 100 200 300 400 500 600

Thre

at S

core

Dust conc. threshold value (μg/m3)

DAY_1

DAY_2

DAY_3

DAY_4

DAY_5

- Statistical verification - Other statistics of dust prediction (MASINGAR mk-2)

90μg/m3

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0 100 200 300 400 500 600

Hit

Rat

e

Dust conc. threshold value (μg/m3)

DAY_1

DAY_2

DAY_3

DAY_4

DAY_5

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0 100 200 300 400 500 600

Fals

e A

larm

Rat

io

Dust conc. threshold value (μg/m3)

DAY_1

DAY_2

DAY_3

DAY_4

DAY_5

0

0.2

0.4

0.6

0.8

1

0

0.2

0.4

0.6

0.8

1

0 100 200 300 400 500 600

Pe

rce

nt

Co

rre

ct

Dust conc. Threshold value (μg/m3)

DAY_1

DAY_2

DAY_3

DAY_4

DAY_5

90μg/m3

90μg/m3

90μg/m3

- Quantitative verification - Predicted dust concentration against surface SPM observation

We use the data that the Ministry of Environment has been operating as the Atmospheric Environmental Regional Observation System called “Soramame-kun” to compare observed surface SPM and predicted dust concentration. We convert the SPM data at each stations into grid point data to match the model grid. Then we calculate time series statistics for each grid.

(Verification period : March-May 2010-2014)

Observed SPM raw data Observed SPM grid data

grid

μg/m3

- Quantitative verification - Predicted dust concentration against surface SPM observation

Statistics MASINGAR MASINGAR mk-2

Mean Error (ME) 20.96 (μg/m3) 3.33 (μg/m3)

Root Mean Squared Error (RMSE)

82.50 (μg/m3)

59.91 (μg/m3)

Correlation Coefficient (CC)

0.45 0.44

All over Japan (Ave. Mar.-May. 2010-2014) • The ME and RMSE are well improved.

• The RMSE is still high and the tendency is remarkable in western Japan.

• We admit a positive bias (ME>0) for dust predictions.

Statistics for each grid map (ME, RMSE, CC)

μg/m3 μg/m3

- Quantitative verification - Case study for predicted dust concentration against surface SPM observation

※Near Fukuoka city (in 2011)

Small dust events

Large dust events

• During small dust events, the current model values show good agreement with observations. On the other hand, the predicted dust concentration is still overestimated during large dust events.

→ As a result, there is a tendency that RMSE is still large. And there is room for improvement in quantitative dust prediction accuracy.

Statistics MASINGAR MASINGAR mk-2

Mean Error (ME) 29.80 (μg/m3) 10.55 (μg/m3)

Root Mean Squared Error (RMSE)

126.53 (μg/m3)

96.91 (μg/m3)

Correlation Coefficient (CC)

0.60 0.55

Near Fukuoka city (Ave. Mar.-May. 2010-2014)

0

500

1000

1500

2000

2500

2011/3/1 2011/3/11 2011/3/21 2011/3/31 2011/4/10 2011/4/20 2011/4/30 2011/5/10 2011/5/20 2011/5/30

Co

nce

ntr

atio

n(μg/

m3)

Date

Observed surface SPM vs. predicted dust concentration (Lat=33.75, Lon=130.00)

MASINGAR

MASINGAR mk-2

SPM

High-resolution global aerosol forecast model

TL159L40 (~110km) TL479L40 (~40km)